tu/e philipsphilips medical systems healthcare it - advanced development 1/38 the effects of...

Post on 20-Dec-2015

215 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

1/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

The Effects of Filtering on Visualization and Detection of Colonic Polyps in Ultra Low Dose Multi-Detector CT Data

Gert SchoonenbergBiomedical Engineering student, TU/e

Final presentation

Supervisors:

Roel Truyen, Anna Vilanova and Frans Gerritsen

Project period:

13-9-2004 – 12-10-2005

2/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Overview

• Motivation: colorectal cancer• Screening methods• Research questions• Dose in Computed Tomography (CT)• Filtering• Computer-aided polyp detection• Visualization• Conclusion• Time for questions

3/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Colorectal cancer

For industrialized countries:• Second leading cause of cancer-related death• Accounts for 10% of all cancer mortality

For the Netherlands:• Causes each year over 9,100 new cases• Causes each year over 4,400 deaths• Accounts for 3% of all deaths

4/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Screening for colorectal cancer

Colorectal cancer:• High prevalence• Long asymptomatic premalignant phase• Well treatable when detected early

Suitable disease

for screening

5/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Outcomes of screening

Target condition

present absent

Diagnostic result

positive TP (true positive) FP (false positive)

negative FN (false negative) TN (true negative)

FNTP

TPysensitivit

FPTN

TNyspecificit

6/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Screening methods

Imaging technique Other technique

Scanner X-ray(DCBE)

Endoscope Proteomics Fecal tests

CT MR Sigmoidoscopy Colonoscopy Blood(FOBT)

DNA

Screening methods

7/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Research questions

• Is computer-aided polyp detection still possible if the dose is reduced?

• Can the artifacts caused by noise in endoluminal visualizations be reduced?

Investigate the use of noise reduction techniques.

8/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Data & CT effective dose

Description Effective dose (mSv)

Comparable to natural background radiation for

CT abdomen (2 scans, 70 mAs

8.4 2.8 years

CT abdomen (2 scans, 6.25 mAs

0.75 3 months

CT abdomen (2 scans, 1.39 mAs

0.17 < 1 month

Mammography 0.7 3 months

Chest X-ray 0.1 10 days

Coast-to-coast round trip in the US

0.03 4 days

9/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Filtering

Gaussian filtering

Bilateral filtering

10/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Gaussian filter

11/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Bilateral filter

12/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results filter methods

Gaussian filteringScale = 2.0 mm

Bilateral filteringScale = 2.95 mmScale = 250 HU

13/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Computer-aided polyp detection

Algorithm developed by Simona Grigorescu and Joost Peters,

Advanced Development, Healthcare IT, Philips Medical Systems Best

14/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Polyp detection overview

Three steps:• Colon segmentation• Polyp detection: identification and detection• Polyp classification

– Bounding box– Linear classifier

Colon segmentation

15/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

PD: identification and detection

• Detection step: shape based– All regions with high curvature are selected– Features are calculated for shell volume– Features are calculated for core volume

airtissue

polyphigh curvature shell

core

colonwall

16/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

PD: Classification

• Feature selection for bounding box– Purpose: discarding outliers.– Select only those features for which the polyp class

is compact.– Select only those features that really discard FP.

Feature 1

Fea

ture

3

Feature 2

Fea

ture

3

Not a polypPolyp

17/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

PD: Classification

• Feature selection for linear classifiers (outliers unwanted)– Rank features according to their Gaussianity

(minimal overlapping).– Forward selection of features which increase the

cluster separability with a minimal value.

bad good

Not a polypPolyp

18/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Training

• Detection step– detect all candidates and calculate shell features

and core features.

• Bounding box– find a minimal cube in the feature space that

contains all polyp examples to get rid of outliers.

• Linear classifier– find a linear boundary between two classes based

on the examples (without outliers).

19/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Testing

• Detection step– Detect all candidates and calculate shell features

and core features.

• Bounding box– Bounding Box: Select only candidates inside the

hypercube.

• Linear classifier– Classification: Classify candidates.

20/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Experiment 1

Test performance of detection step using:

• Bilateral filtered data• Unfiltered data

All parameters in the CAD algorithm are kept constant.

21/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: detection step

Bilateral filtering 33% FP reduction 2.2% decrease in sensitivity (polyps > 6 mm)

Dose level Filtering Sensitivity FP

Normal None 95% 134

Low None 95% 163

Ultra low None 95% 312

Normal Bilateral 93% 97

Low Bilateral 92% 109

Ultra low Bilateral 93% 201

22/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Feature selection for classification

Experiment 2a:• Optimal feature set for each dose level

Experiment 2b:• Optimal feature set for normal dose trained on

normal dose and tested on all dose levels

Experiment 2c:• Robust feature set trained on normal dose and

tested on all dose levels

23/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results 2a: optimal features

Bounding Box

Average false positive reduction is 67%

No loss in sensitivity for polyps 6 mm and larger

In parenthesis () the results of the detection step are given.

Dose level Filtering Sensitivity FP

Normal None 100% (95%) 50 (134)

Low None 100% (95%) 60 (163)

Ultra low None 100% (95%) 71 (312)

Normal Bilateral 100% (93%) 40 (97)

Low Bilateral 100% (92%) 28 (109)

Ultra low Bilateral 100% (93%) 66 (201)

24/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: linear classifier

25/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: linear classifier

26/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results 2b: Normal dose features

Bounding boxOptimal feature set for normal dose trained on

normal dose and tested on all dose levels.Decrease of sensitivity on lower dose levels!

In parenthesis () the results of the detection step are given.

Dose level Filtering Sensitivity FP

Normal None 100% (95%) 50 (134)

Low None 56% (95%) 50 (163)

Ultra low None 12% (95%) 44 (312)

Normal Bilateral 100% (93%) 40 (97)

Low Bilateral 85% (92%) 38 (109)

Ultra low Bilateral 9% (93%) 28 (201)

27/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Result: linear classifier

28/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results 2c: Robust features

Bounding boxNot filtered data: Sensitivity: 96%, average FP reduction: 30%

Bilateral filtered data: Sensitivity: 98%, average FP reduction: 40%

In parenthesis () the results of the detection step are given.

Dose level Filtering Sensitivity FP

Normal None 100% (95%) 50 (134)

Low None 96% (95%) 50 (163)

Ultra low None 93% (95%) 44 (312)

Normal Bilateral 100% (93%) 40 (97)

Low Bilateral 100% (92%) 38 (109)

Ultra low Bilateral 95% (93%) 28 (201)

29/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: linear classifier

30/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Visualization

31/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Perspective ray-casting

32/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Current visualization

Normal doseSmooth surface

Low doseBlobs appear

Normal doseRough surface

33/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Proposed solutions

• Bilateral filtering blobs• Gradient smoothing rough surface

34/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: normal dose

35/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Results: all dose levels

1.6 mAs 6.25 mAs 64 mAs

36/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Conclusions

Very low dose: filtering before rendering is required!

V The new rendering (gradient smoothing) gives similar renderings for low and high dose.

V With the new rendering the wall appears relatively smooth when it in fact should be smooth.

X No smoothing of important structures.The noise level changes within a dataset. In the really noisy regions strong filtering is needed and smoothing occurs. [New scanners: dose modulation]

37/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Overall conclusion

• Visualization and computer-aided detection of colorectal polyps is feasible on ultra low dose CT colonography data.

• It is also possible to create one visualization algorithm and one computer-aided detection algorithm that can cope with various dose levels.

38/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Time for questions

Questions?

39/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Current visualization

Problem:Bumpy colonsurface.

Cause:Not theisosurfacelocation, butsurface normals.

Data:64 mAs datapelvic region

40/38

TU/e

PHILIPS Philips Medical SystemsHealthcare IT - Advanced Development

Simulating low dose

01

1

1 1

nqn poidev

q

q: ratio of actual and desired mAs level

poidev: Poisson distribution

n0: detected photons

n1: simulated amount of photons

top related